Saturday, January 18, 2014

Three Roads to Artificial Intelligence

AI pioneer Marvin Minsky once famously stated that “Artificial Intelligence is the science of making machines do things that would require intelligence if done by men.”

Such an elegantly recursive cop-out!

I’m going to propose three, more operational models of intelligence: competence, search and creativity respectively. Which seems more plausible to you?

1. Competence

You meet an expert and watch them work at solving problems. A query comes in and instantly the correct answer rings out. You recall the old saying: “An expert is someone who doesn't have to think, because they know.”

This kind of intelligence is algorithmic. In principle you can write a program simulating the expert which delivers answers in a computationally well-behaved fashion (code for polynomial behaviour). In AI we call these programs Expert Systems.

This is intelligence-as-instinct: compiled, hard-wired expertise. I've met people like this and some I wouldn't call smart at all – the ones who are flummoxed by an unfamiliar problem.

2. Search

Some people have equated intelligence to controlled search. The paradigmatic example is playing games such as chess where there is no known well-behaved algorithm which can take an arbitrary game position and return the optimal next move. The best games programs create a look ahead tree using legal moves and assess the best of the future game-states. They then choose their next move as best-placed to get to that state despite the best attempts of their opponent.

A chess lookahead tree

Search can look quite intelligent because it’s flexible and adaptive. The AI program doesn’t know what you’re going to do next, but whatever you do it will adapt and continue “intelligently” towards its final goal of winning. I have the same feeling about my sat nav, which implacably directs me to my destination no matter what wrong turns and detours I make.

Search is powerful and adaptable (trading competence for bounds on space and time resources) but suffers from a fatal rigidity: it explores just the possibilities defined by the state-space and operations given to it. The chess program just plays chess.

Important as the distinction is in artificial intelligence, it’s not clear to me that in humans, search and competence are that much different. Humans are very bad at search, finding it almost impossible to hold a large number of possible future states simultaneously in mind. Trying to solve problems in such a way is pretty much the definition – in humans – of incompetence.

Experts differ from novices not by doing more search, but by having a more extensive, refined and sophisticated competency set (or ‘knowledge base’).

3. Creativity

The people I find truly, scarily intelligent are those who keep you off-balance by continually moving the goalposts in a way both surprising and opaque. You feel your every possible gambit has already been anticipated and that the activity you think you're conducting is actually embedded in a much more complex scenario being deftly manipulated by your opponent.

In “Tactics of Mistake’, Gordon R. Dickson describes an enemy force advancing down a river valley along its narrow flood plain; the friendly force is much smaller. A merely competent commander would presumably choose the best combat tactics commensurate with his poor hand – and would expect to lose. In fact Dickson’s hero places his troops in well-screened locations in the hills to the side of the valley, and then dams the river. As the water level rises, the flood plain floods and the enemy troops are forced into a killing zone. They are thus defeated.

How is this solution-approach different from competence and search?

The critical factor is that new elements have been brought in to create a larger ‘game’ – in particular the river, its flooding behaviour, the typography of the ground and the possibilities of damming.

Creativity thus requires an additional context, extra resources which can be brought to bear on the original goals of the game. In the real world, everything we do is embedded in layers of enveloping reality. For example, you might defeat the chess champion by doping his coffee so that he plays particularly poorly. You might thus win in the extended game where the player is also an active constituent, but a chess program has no access to this larger reality. More legitimate 'psychological tricks’ are regularly employed by human players.

This points to an important feature of creative intelligence. It requires a deep familiarity with the potential of embedding contexts of the proximate ‘game’ or problem – and the ability to select and refine additional operators which can be played back into a new kind of solution.

Most games are defined to explicitly abstract away all inessential contexts: you are allowed to do just what the rules say and no more (so no stealing your opponent’s king and declaring victory!). But this immediately rules out the kind of creativity we're discussing here, thereby impoverishing the model of intelligence which can be studied or deployed. I’m not sure AI research has sufficiently taken this into account.

In the real world there are no absolutely impermeable boundaries, so there are potentially no limits to the kinds of esoteric knowledge which can be brought to bear on a problem. And that’s what truly intelligent people do.

What is measured by IQ tests?

Test of crystallized intelligence (such as general knowledge) seem to be measuring competence which they hope correlates with ‘g’ as a proxy. However, the core of intelligence seems to be fluid intelligence, measured by test items such as Raven’s Progressive Matrices. These require the inference of new, compelling rules and patterns from presented data - which sounds a lot more like creative intelligence.

Let me make one more remark - about scientists and mathematicians. Some people are very good at rapidly and easily seeing the consequences of assumptions; they sign-up to the "shut up and calculate" school. Others seem more comfortable exploring different paradigms for situating a problem, other ways of thinking about it. Theoretical physicist Lee Smolin called these two types "craftspeople and seers", while Freeman Dyson preferred "Frogs and Birds". Perhaps there is a connection here with the intelligence-as-search and intelligence-as-creativity distinction?

Psychologist Daniel Nettle observed that intelligence is a kind of whole-brain efficiency measure implicated across all areas of neural functioning.  High-scorers on the personality trait of Openness are artistic, creative people capable of making associations between different – and perhaps surprising – kinds of things (Smolin's "seers") while those with a more "craftsperson" style are perhaps exhibiting the effects of high IQ per se.